Generation of peptide detectability datasets from single DIA experiment for prediction model fine-tuning

Area of Science:

• Proteomics
• Mass Spectrometry
• Computational Biology

Background:

• Accurate peptide detectability prediction is crucial for mass spectrometry-based proteomics, impacting protein identification and quantification.
• Existing sequence-based models have limited applicability due to instrument and experimental condition variability.
• Current fine-tuning methods require extensive datasets (up to 300,000 peptides), incurring high costs.

Purpose of the Study:

• To present a novel approach for generating peptide detectability datasets from a single Data-Independent Acquisition (DIA) experiment.
• To enable efficient fine-tuning of prediction models with minimal raw data for specific experimental conditions.
• To reduce the substantial data and cost requirements associated with traditional model training.

Main Methods:

• Developed a complementary method to generate peptide detectability datasets directly from a single DIA experiment.
• Utilized these datasets to fine-tune peptide detectability prediction models.
• Applied predicted detectability for filtering search libraries.

Main Results:

• The new approach enables model fine-tuning with minimal raw data, improving adaptation to experimental conditions.
• This strategy significantly reduces data and cost requirements for training prediction models.
• Filtering search libraries using predicted detectability led to increased peptide identification rates and decreased computational time.

Conclusions:

• The presented method offers a cost-effective and efficient way to generate peptide detectability datasets.
• This approach enhances the adaptability and accuracy of prediction models in mass spectrometry.
• Optimizing search libraries with predicted detectability improves proteomics experimental outcomes.